Tokens are becoming the operating unit of language models - and increasingly, the currency of AI agents.
They measure capacity and shape cost: how much context a model can hold, how much evidence it can reason over, and how much work it can produce in one run.
Input, output, cached context, long context - each now has its own economics. What once looked like a technical detail is becoming a practical constraint in how AI systems are designed, priced, and used.[1-4]
So how should we translate the abstract numbers - 128k, 1M, 2M - into real work?
This series maps token economics onto professional artifacts: clinical trial protocols, regulatory documents, drug-discovery literature, financial filings, analyst reports, contracts, and case law.
Not just how many words fit.
But what evidence fits.
What needs retrieval.
What becomes expensive.
What becomes possible.
128k
may hold a full protocol, a long financial filing, or a substantial contract set.
1M
can begin to hold curated evidence packs: dozens of papers, multiple filings, litigation materials, or due-diligence documents.
2M
moves closer to archive-scale work - but still does not remove the need for selection, structure, and validation.[5]
The practical question is no longer just: how large is the context window?
It is: what kind of work does that window actually make possible?
A context window is not only a model specification.
It is a working memory budget.